The latest public release of facebookresearch/xformers is v0.0.35. This repository is a Python toolbox collecting optimized building blocks that can be used for Transformer research. The provided record contains no changes for this version, so what was added or fixed in v0.0.35 cannot be confirmed.
xFormers provides independent, user-customizable components that can be combined and used in vision, NLP and other areas. It focuses on making use of research components that are not yet included in mainstream libraries such as PyTorch. It ships with its own CUDA kernels and dispatches operations to other libraries when needed. What it offers includes memory-efficient exact attention and sparse attention, as well as fused operations that combine several operations. So rather than a single finished model, it is closer to a parts kit that researchers use to assemble and experiment with Transformer operations.
What can be verified extends only as far as the repository description and the features listed in the README. The provided material contains no specific change log or performance comparison for v0.0.35, so it is impossible to judge what this release improved over previous versions. The license entry is also marked NOASSERTION, so usage conditions cannot be settled from this record alone. It is therefore difficult to assert anything about the novelty of this release or its usage conditions without separate evidence.